Research And Application On Ship Detection Algorithm Based On YOLOv7 And Attention Mechanism

The water transportation system is a key research object in China’s transportation field,and the research on ship object detection is of great significance for maintaining water transportation safety,exploration of marine resources,disaster relief,and national defense security in China.However,traditional object detection methods are difficult to accurately detect ship objects in real time under the conditions of insufficient light,ship obstruction,and background interference and so on.In recent years,with the rapid development of deep learning image recognition technology,intelligent video surveillance based on deep learning has been a research hotspot in the field of maritime supervision and service.Deep learning can effectively extract features on multi-scale images and adaptively process redundant information,which provides favorable assistance for accurate detection of objects.Based on this,this dissertation takes ship images as the research object,uses deep learning methods as the basic theory,designs the deep learning model structure,optimizes the model’s feature extraction ability,and lightens the model.The specific research content of this dissertation is as follows:(1)A ship classification detection method based on YOLOv7 and residual convolutional block attention module(YOLOv7-RCBAM)is proposed to address the problems of inaccurate target feature extraction and unobvious deep feature information.First,to speed up the training,transfer learning is used to freeze the backbone network parameters of the pre training model,and fine tune the model to train the downstream network.Second,to enhance the information correlation of channel dimension features,a channel attention mechanism with residual connections is proposed.Final,a feature fusion attention mechanism method is proposed to improve the effectiveness of object features.After experimental verification on the Seaships dataset,the accuracy of the proposed method reached 97.59%,indicating that the proposed method effectively improves the model’s ability to extract deep information;Meanwhile,in the experiment of image enhanced ship dataset,the detection accuracy of the proposed method reached 96.13%,indicating that the model has strong anti-interference ability.(2)A lightweight Swin-YOLOFormer ship classification detection method is proposed to address the problem of slow model detection speed and difficulty fitting datasets due to the large volume and parameter redundancy of the YOLOv7 model.First,in terms of the backbone network,the Swin Transformer lightweight model has been introduced,effectively reducing the redundancy of the backbone network parameters.Second,on the feature fusion network,an improved G-ELAN-H module is proposed to extract features and optimize the ELAN-H convolution module to reduce the computational burden of model parameters.Final,to prevent feature loss,an improved SPPCPSC module is proposed to enhance the features of the Receptive field.Through experimental verification on a self-made ship dataset,compared with the benchmark model,the proposed model parameter quantity decreased by 66.05% to13.501 M,not only accelerating the model training speed,but also achieving an accuracy of97.81%.The experiment shows that the proposed method achieves model lightweight.(3)Based on the previous research,a lightweight ship detection application system based on the Swin-YOLOFormer model is designed.The system supports functions such as model loading,ship image detection,and video detection.The application system test results show that the application has high detection accuracy,accurate target positioning,and fast image detection response.The experiment has proven the practicality of the Swin-YOLOFormer ship detection application system designed.